DeepSeek-R1 — Open-Weight Reasoning Model Rivaling OpenAI o1

Quick Use

1. Hosted: same DeepSeek API key, set model="deepseek-reasoner"
2. Local: ollama pull deepseek-r1:7b && ollama run deepseek-r1:7b
3. Print response.message.reasoning_content to see the full chain-of-thought

Intro

DeepSeek-R1 is the open-weight reasoning model that achieves o1-level performance on AIME / MATH / GPQA / Codeforces while shipping its full chain-of-thought to the user. Distilled smaller versions (1.5B, 7B, 32B, 70B) make local reasoning practical on consumer hardware. MIT license, full weights public. Best for: hard reasoning tasks (math, science, complex code) where you need a reasoning model but want open weights. Works with: DeepSeek API, Ollama (distilled), vLLM, llama.cpp. Setup time: 2 minutes.

Hosted API

from openai import OpenAI

client = OpenAI(
    base_url="https://api.deepseek.com/v1",
    api_key=os.environ["DEEPSEEK_API_KEY"],
)

response = client.chat.completions.create(
    model="deepseek-reasoner",  # R1
    messages=[{"role": "user", "content":
        "Prove that the square root of 2 is irrational"}],
)

# R1 streams reasoning + final answer
for choice in response.choices:
    print("REASONING:", choice.message.reasoning_content)
    print("ANSWER:", choice.message.content)

Unlike o1, R1's reasoning is visible — useful for debugging, education, and trust.

Local via Ollama (distilled)

ollama pull deepseek-r1:1.5b   # ~1GB, runs on a laptop
ollama pull deepseek-r1:7b     # ~5GB
ollama pull deepseek-r1:14b    # ~9GB
ollama pull deepseek-r1:32b    # ~20GB, M2 Max territory
ollama pull deepseek-r1:70b    # ~40GB, beefy server

The 7B distillation often outperforms GPT-4o on competition math while being free and fast on a single 4090.

When to use R1 vs V3

Pricing

FAQ

Q: Why does R1 show its reasoning when o1 hides it? A: DeepSeek published the full RL training methodology. Visible CoT is part of the value proposition — auditability, debugging, education. OpenAI considers o1's CoT a competitive moat.

Q: How much slower is R1 vs V3? A: R1 spends extra tokens on reasoning before the final answer — typically 3-10× more output tokens, so 3-10× slower wall-clock latency on equal infra. The cost difference reflects this.

Q: Are the distilled R1 versions trained from scratch? A: No — they're knowledge-distilled from full R1 into Llama / Qwen base models. The 7B distill is Llama-3.1-8B + R1 distillation, the 32B is Qwen-2.5-32B + R1 distillation, etc. Performance trades off with base.

Source & Thanks

Built by DeepSeek. Weights MIT-licensed.

deepseek-ai/DeepSeek-R1 — ⭐ 90,000+